Abstract
<jats:p>Bolted flange joints are critical components in offshore and subsea systems, where loss of bolt preload can lead to leakage, environmental damage, and costly operational downtime. Smart touch–based stress wave sensing using piezoceramic transducers has emerged as an effective non-intrusive technique for bolt looseness detection; however, most data-driven approaches rely on large amounts of labeled underwater data, which are expensive and difficult to obtain in practice. Moreover, stress wave signals collected in underwater environments exhibit substantially different characteristics from those measured in air, resulting in severe domain shift that limits the direct transfer of models trained in laboratory conditions in air. To address these challenges, this paper proposes a few-shot air-to-water domain adaptation framework that combines a Domain-Adversarial Neural Network (DANN) with a Centroid Alignment Domain Adaptation (CADA) strategy. Mel-frequency cepstral coefficients (MFCCs) are first extracted from stress-wave responses generated by a smart touch excitation mechanism. DANN is employed to learn domain-invariant and label-discriminative feature representations, while CADA explicitly compensates for class-wise distribution mismatch by aligning source-domain feature centroids with those estimated from a small set of labeled underwater samples. The aligned features are subsequently used to train a support vector machine for torque classification. The proposed framework is experimentally validated on bolted flange assemblies with diameters of 5, 6, and 9 inches under air-to-water transfer scenarios. Results demonstrate that the proposed DANN–CADA approach consistently outperforms a baseline DANN model across all flange sizes, achieving significant improvements in classification accuracy while using only a few labeled underwater samples. These findings indicate that the proposed method effectively mitigates both covariate and concept shifts between air and water environments. By substantially reducing the need for extensive underwater data collection, the proposed framework offers a practical and scalable solution for underwater bolt looseness detection in offshore and subsea structural health monitoring applications.</jats:p>